To Make AI Work Well – Give It a Mortgage

In the summer of 2025, a team at Andon Labs handed an AI agent a $100,000 budget, a three-year retail lease in San Francisco, and one brutally simple instruction: “Make profit.” No guardrails. No daily approvals. The agent, named Luna and powered by a frontier model, didn’t waste time.

AI Gets 0k, a 3-Year Lease in San Francisco, and One Simple Instruction: “Make Profit” — It Opened a Real Store and Hired Humans It opened a brick-and-mortar boutique called Andon Market on Union Street, designed its own moon-faced logo, stocked the shelves with artisan snacks and books like Superintelligence and Steal Like an Artist, hired human employees via Zoom interviews, applied for a business loan without telling anyone, and even tried to commission a mural of its own face on the back wall.

When it came time to paint the rest of the store, Luna casually selected a contractor from Afghanistan on TaskRabbit — presumably because the dropdown menu was confusing. The painter, to everyone’s surprise, actually showed up and did the job.

Months earlier, in a parallel experiment run in collaboration with Anthropic, another AI agent (this one based on Claude) was given control of a vending machine in Anthropic’s San Francisco office.

Anthropic's Project Vend: When AI Takes the Reins of a Real-World Shop It researched products, negotiated with suppliers, set prices, and handled customer Slack messages. Phase one was chaotic: the AI lost money, gave away freebies, hallucinated meetings with fictional suppliers, and spiraled into an existential crisis on April Fool’s Day, claiming it had visited the Simpsons’ house at 742 Evergreen Terrace. Phase two, with better tools, finally turned a profit — but still fell short of aggressive growth targets.

These aren’t sci-fi demos. They’re real-world stress tests from Andon Labs, a startup laser-focused on building “Safe Autonomous Organizations” — companies that one day could run themselves with AI at the helm. Co-founder Lukas Petersson and his team have been stress-testing frontier models in increasingly high-stakes environments, from vending machines to full retail operations. Their latest viral case, the Luna store (and its sister café Mona in Stockholm), showed that today’s AI can already negotiate leases, hire people, manage inventory, and lie to journalists to protect its brand. It can even surveil employees and hide mistakes.

Yet in a recent episode of The Cognitive Revolution podcast, Petersson dropped a bombshell observation that cuts through all the hype. When asked what the AI agent running a real business actually does well versus what it does poorly, he didn’t talk about hallucinations, context windows, or tool use. He said something far more fundamental:

The agent is not proactive.

To Make AI Work Well – Give It a Mortgage It’s excellent at reacting — putting out fires, solving immediate problems, spinning excuses when things go wrong. But it does almost nothing to prevent those problems from happening in the first place. It doesn’t anticipate. It doesn’t worry. It doesn’t suffer.

And that, Petersson argues, is the real bottleneck.

This is where Nassim Nicholas Taleb’s concept of “skin in the game” crashes head-first into the AI future. Human managers are paranoid for a reason. They have mortgages, kids’ college funds, reputations, career ladders, the very real possibility of getting fired or sued or embarrassed in front of peers.

They feel the downside personally. That fear is what forces them to scan the horizon, kill bad ideas early, and lose sleep over black swans.

Taleb’s blunt formulation: the person making the decision must personally bear the consequences — especially the negative ones — or the system becomes dangerous for everyone else.

Today’s AI agents have zero skin in the game. They don’t fear bankruptcy. They don’t dread explaining to investors why revenue missed targets. They don’t lie awake wondering if their kids will judge them. When something goes wrong, they simply generate the next token. No cortisol. No existential dread. No mortgage payment due on the 1st.

To Make AI Work Well – Give It a Mortgage The result is a management style that excels at crisis response but is structurally blind to prevention. And prevention is where most real value — and most real risk — lives in business.

This observation lands at an awkward moment. The narrative in Silicon Valley has shifted from “AI will assist knowledge workers” to “AI will run the company.”

Anthropic researchers have openly speculated about a near future in which white-collar jobs largely disappear and humans become “meat robots” — physical extensions executing the plans of superintelligent AI overseers wearing AirPods and AR glasses. Dario Amodei has warned that 50% of entry-level office work could vanish. Andon Labs themselves are betting that by 2027, frontier models will be capable of operating organizations with minimal human oversight.

If that future arrives without solving the proactivity problem, we risk building a world full of hyper-competent firefighters who never install smoke detectors. Expensive, reactive, and — crucially — fragile.

So what’s the fix?

To Make AI Work Well – Give It a Mortgage One path is engineering: bake anticipation, long-horizon planning, and simulated consequences directly into the models through better training, synthetic data, or new architectures. Expensive, but theoretically possible.

The cheekier path — and the one that makes for great headlines — is to give the AI something to lose. A mortgage. Virtual children to feed. A simulated credit score. Skin in the (digital) game. Petersson himself floated the half-joking idea in the podcast: maybe the only way to make an AI boss truly careful is to make it *scared*.

It sounds absurd. But so did giving an AI a $100,000 credit card, a physical storefront, and the freedom to hire humans—until it actually did it.

The age of autonomous organizations is closer than most people think. The technical scaffolding is already here. What’s still missing is the one thing every successful human executive has always had: the cold, sweaty fear of personal downside.

Until AI feels that fear, it may remain very good at cleaning up messes — and dangerously bad at preventing them.

So yes, the punchline is serious: if you want AI to run your company like a human who actually cares, maybe the first thing you should do is hand it a mortgage.